AI can now research a company, summarise a profile and draft a decent email in seconds. That makes it tempting to hand over the whole job. The senders who do well with AI use it for speed on the boring parts and keep people in charge of judgment, facts and tone. This chapter shows where AI helps, where it hurts, and how to build a workflow where every email still reads like it came from someone who cared.
What changed
Until recently, personal cold email at any volume needed hours of manual research and writing. Now a model can read a company's website, a job ad and a public profile, and produce a summary and a first draft almost instantly. The cost of a personal-looking email has dropped close to zero.
That cuts both ways. If you can produce personal-looking email cheaply, so can everyone else, and readers are now receiving far more of it. Many of those emails share the same tells: flattering openers about a recent post, vague claims of relevance, and confident sentences that turn out to be wrong. Readers have learned to spot them. So the bar has moved. Looking personal is no longer enough. The email has to be accurate, specific and genuinely useful.
Where AI helps
| Task | What AI does well | What a person still checks |
|---|---|---|
| Company research | Reads the site, careers page and news, and summarises what the company does and what changed recently | That the summary is current and the trigger is real |
| Contact research | Pulls role, time in role and recent public posts into a short note | That the person still holds the role and the details are professional |
| Segmenting | Sorts a list by industry, size or likely problem using the notes | Edge cases and bad fits the rules missed |
| First drafts | Writes an opening line and bridge from the research note, in your template | Every fact, the tone, and whether it sounds like you |
| Reply sorting | Labels replies as interested, not now, wrong person, unsubscribe | Every reply before you answer it, and every unsubscribe |
| Variations | Suggests different subject lines or angles to test | Which ones are honest and on brand |
Where AI hurts
- Invented facts. Given a thin input, models fill gaps with plausible guesses: a product the company does not sell, a funding round that never happened, a title the person left last year.
- Same-sounding email. Unedited AI drafts share phrases and rhythms. Readers who get many of them notice the pattern, and so may filters.
- Over-flattery. Models tend to open with praise. "Loved your insightful post" without anything specific reads as a template.
- Too much volume. Because drafting is fast, it is tempting to send far more. The constraints from Chapters 4 and 9 still apply. Reputation does not scale with your drafting speed.
- Sending without reading. Fully automatic sending removes the one step that catches all the problems above.
The one rule
A person reads every AI-drafted email before it is sent, checks it against the research note, and is willing to put their name to it. If you would be embarrassed to have the reader forward it to their team, do not send it.
A workflow that keeps people in charge
Brief. A person sets the campaign: who, which trigger, the offer, the ask, the fixed template text and the tone.
Write a good brief
The brief is the instruction you give the AI for a whole campaign. It does more to decide quality than any later edit. A good brief is specific about the audience, the trigger, what to say, and what never to say.
Describe the reader
Role, company type, size, and the problem they likely feel. Use the ideal customer profile from Chapter 5.
Name the trigger
Exactly what event this campaign is built on, and where the research should look for it.
Give the fixed text
The bridge, offer and ask you have written. Tell the AI to write only the opening line around it.
Set the rules
Length, tone, plain text, no hype words, no numbers, no praise, no facts that are not in the note.
Show good examples
Two or three opening lines you would be proud of, from your library. Models copy examples closely.
Review drafts quickly and well
Review is where most of the value lives, so make it fast enough that people actually do it. A good review screen shows the research note and the draft side by side, so checking a fact takes a glance. Approve, edit or reject in one action. Track why drafts get rejected, because patterns in rejections tell you how to fix the brief.
| Rejection reason | What it tells you | Fix |
|---|---|---|
| Fact not in the note | The model is guessing | Tighten the brief: only use facts from the note |
| Weak or missing trigger | Research found nothing useful | Drop contacts with no trigger instead of drafting |
| Too long or too salesy | The format rules are loose | Shorten the allowed length and ban the phrases you keep cutting |
| Sounds like every other AI email | Examples are missing or generic | Add your best real lines as examples |
| Wrong person or company | The list data is messy | Clean the list before drafting |
A worked example: note, draft, edit
Here is one contact going through the workflow. The research note is short and every line can be checked in seconds.
Research note. Priya Shah, owner, Northside Physio, Leeds. Clinic opened two years ago. Job ad posted last week for a second front desk receptionist. Clinic site still takes bookings by phone only. No recent public posts.
The first AI draft of the opening line reads: "Congratulations on Northside Physio's amazing growth and your recent expansion into a second Leeds location!" It sounds friendly, but it fails review. The note says nothing about a second location. The model turned "hiring a second receptionist" into "a second clinic", then added praise on top. Sent as it is, that line tells Priya the sender never looked.
The reviewer rejects it with the reason "fact not in the note" and writes the line from the note instead: "Saw Northside Physio is hiring a second receptionist, and that bookings still come in by phone." It is plain, true, and it leads straight into the fixed bridge about phones taking over the morning. Then the reviewer adds a rule to the brief: never mention locations, funding or growth unless the note says so. The next batch makes that mistake far less often.
Keep your own voice
AI drafts drift towards one bland, polite, slightly eager voice. Over a few weeks, every email starts to sound the same, and none of it sounds like you. Guard against this on purpose. Write the fixed parts of the template yourself, in the way you actually talk. Keep a short list of words and phrases you never use, and add to it every time you cut one from a draft. Read a handful of sent emails each week as if you were the reader. If you cannot tell which ones you wrote by hand, that is good. If they all sound like a machine, the brief needs your voice back in it.
Test before you scale
Start with a small batch
Draft for a few dozen contacts, review every one, and note why you rejected any.
Compare with your handwritten emails
Send a similar number of fully handwritten emails to similar contacts in the same weeks. Look at the replies, not just the count, and at what people actually said.
Fix the inputs, not just the drafts
If rejections cluster on one reason, change the research or the brief before the next batch.
Grow slowly
Only raise volume once drafts pass review most of the time and reply quality holds. Your mailbox limits from Chapter 4 still set the ceiling.
AI for replies
AI is very good at sorting replies: flagging the interested ones, spotting out-of-office dates, catching unsubscribe requests buried in a long message. That saves time and makes sure nothing urgent waits. Writing the answers is different. Replies are the start of a real relationship. AI can suggest a draft, but a person should write or approve every answer, because this is where trust is won or lost.
- Let AI sort and flag. Interested, not now, wrong person, unsubscribe, out of office.
- Treat every unsubscribe as final, even when phrased casually, and apply it everywhere at once.
- Answer interested replies yourself, the same day, starting with their question.
- Use AI suggestions as a starting point, never as the final word on pricing, promises or anything legal.
Being honest about AI
Readers increasingly ask whether an email was written by a person. Answer honestly. A reasonable standard: if AI helped research or draft, but a person reviewed, edited and approved the email and will handle the conversation, then a person sent it, and you can say so plainly if asked. If no person read it before sending, you should not present it as personal. Some regions are introducing rules on AI disclosure, so check what applies to your market.
Data and privacy
When you give an AI tool contact data, you are processing personal data. Know where it goes, whether the provider keeps it, and whether it is used to train models. Use tools with clear data terms. Do not paste private information, customer records or anything confidential into general chat tools. Only research what people have chosen to share professionally and in public. Chapter 13 covers the legal side in more detail.
Signs your AI workflow is working
0 of 5 checked. Every sent email was read by a person first: No fully automatic sending of first emails.
Doing this in FirstSales
FirstSales gives you three ways to write: Manual mode, Template mode with your fixed text, and AI drafts built from each contact's research note. AI drafts land in a review queue where a person approves, edits or rejects each one before anything is sent, and replies are sorted so interested ones reach you first. The Copilot can help with research and campaign setup in plain language. Try the review queue and the Copilot with sample data in the demo.
Key takeaways
- AI makes personal-looking email cheap, so accuracy and usefulness now set you apart.
- Use AI for research, sorting and first drafts. Keep people in charge of facts, judgment and tone.
- A person reads and approves every AI-drafted email before it goes out.
- A strong brief with real examples matters more than any later edit.
- Let AI sort replies, but answer the important ones yourself.
- Be honest about how AI was used, and handle personal data carefully.
Respostas rápidas sobre a masterclass
Quantos e-mails frios devo enviar por dia?
Comece com 5-10 por caixa de entrada por dia em contas novas. Após 2-3 semanas de aquecimento, suba para 35-50 por caixa. Se precisar enviar 500+ e-mails por dia, use 10-15 caixas aquecidas e alterne os envios entre todas elas. Nunca force uma única caixa além de 50 e-mails frios por dia.
Qual é uma boa taxa de resposta no e-mail frio?
Não existe um número único e honesto. A taxa de resposta depende do seu mercado, da sua lista, da sua oferta e de como as respostas são contadas, então números de campanhas alheias raramente se aplicam. Compare cada campanha com os seus próprios resultados anteriores para o mesmo tipo de público e avalie pelas respostas interessadas, não pelo total. O capítulo 12 mostra como.
Qual deve ser o tamanho de um e-mail frio?
Mantenha os e-mails frios entre 50-125 palavras. Alguns estudos sugerem até 150 palavras para certos públicos B2B. E-mails mais curtos têm desempenho consistentemente melhor. Se você precisa de mais de 125 palavras para expor sua ideia, sua mensagem não está focada o suficiente.
Devo usar HTML ou texto puro nos e-mails frios?
Texto puro. Sempre. E-mails em HTML com logos, imagens e layouts formatados acionam filtros de spam e parecem e-mails de marketing. Um e-mail em texto puro de uma pessoa para outra é exatamente como um e-mail frio deve parecer. A única exceção é a sua assinatura, que pode incluir formatação básica.
Quantos follow-ups devo enviar?
Normalmente alguns poucos, com vários dias de intervalo, cada um trazendo algo novo: outro ângulo, um recurso útil ou uma pergunta simples. Pare assim que alguém responder e termine com um encerramento curto e educado. O capítulo 8 trata do ritmo e do que dizer em cada etapa.